Hello, I'm Maneshwar. I'm building git-lrc, a Micro AI code reviewer that runs on every commit. It is free and source-available on Github. Star git...
For further actions, you may consider blocking this person and/or reporting abuse
Having "morse code" as your title is crazy. I got scared seeing it lol. btw, great post!
Lmao, thanks bud!
damn what a title.. imagine google ranking it, would be funny
hehe yo xD
that binary breaks faster than people expect. whether Claude truly reasons is less useful than whether you can rely on the output for a specific problem class. for dot puzzles: apparently yes.
xD
haha - somewhere around the third retry it kind of stops being a philosophy question
Hehe xD
right? by retry 3 it's just acceptance. commit the workaround, move on.
I think this falls into the lines of how LLMs operate. Tensors arent there to find tokens, tokens are applied afterwards to the pattern. The pattern is what the LLM recognizes, the tokenizer just puts it into words, so it's useful (think of it as translation layer).
I see
The compression framing is the one that makes this click. Once you picture training as "squeeze all this text into the smallest model that still predicts it," general skills like counting and mirroring stop looking magical and start looking necessary. The one spot I'd soften is "a puzzle it's never seen before," since rising-then-falling symmetry and palindromes are everywhere in the training data, so the structure itself is very familiar even if your exact dots aren't. That actually strengthens your point, because it shows the model reusing a learned operation rather than needing to have seen your specific string.
Predicting the next token is the goal it was trained on, not how it gets there. To do it on inputs it never saw, it has to actually work things out. And not just in the visible reasoning steps - even producing a single token, there's a multi-step process running inside first. A kind of emergent reasoning in the latent space, before any text comes out. Interpretability work has traced it: the model tries several rough approaches in parallel, like probes testing different routes, then combines them. People have watched it add two numbers this way - one path ballparks the sum, another locks in the last digit, and they merge into the answer.
The dot puzzle is the weakest way to make this case, though. Any single clean answer can be waved off as "it saw something close in training," and palindromes are everywhere in text, so that escape hatch is wide open - which is the out a couple of your commenters already took. The internal traces are better evidence because you can watch the work happen no matter what the model saw before.
I believe that the model has already learned some of the dotted annotations like the one which you have explained. We live in a small world there the thoughts overlap 😁
Hehe, true Ranjan :)